@duog, did you disable CPU frequency scaling and otherwise quiet down background tasks when taking these measurements? The changes in CPU times seem a tad unexpected. In particular, I wouldn't have thought that mutator time would change appreciably.

I didn't disable CPU frequency scaling, or try too hard to quiet down background tasks. I did use cset shield, giving 8 cores to the benchmarks (i.e. exclusive cpuset), which I thought would be good enough?

Mutator -15% sure is surprising.

There is a small problem with the patch, causing one of the condition variables to be signalled far too often which I'm fixing.

Reran here https://gitlab.haskell.org/-/snippets/1867

similar (though worse a bit worse) results.